Mistral Small 4

Claims and evidence

• Released 16 March 2026 ● Primary source[1] [4]March 16 (2026) [3]page header date

• Status Available ● Primary source[3]GA badge [5]Generalist models [2]

• Successor of Mistral Small 3.2 EditorialEditorial link along the Mistral Small line: the announcement calls Small 4 the next major release in the Mistral Small family and compares its non-reasoning mode with the chat style of Mistral Small 3.2, the most recent Small release before it, without naming a predecessor; the models overview names Small 4 as the alternative for the retired Small 3.2. ● Primary source[1]first paragraph; Reasoning on demand [5]Deprecated & retired models table, Alternative column

• Change · Reasoning Adds reasoning effort set per request (none or high) in the same model, instead of a separate Magistral reasoning model.Compared with Mistral Small 3.2 ● Primary source[1]Reasoning on demand
• Change · Architecture Moves to a mixture-of-experts design with 128 experts, 4 active per token, and 119B total parameters.Compared with Mistral Small 3.2 ● Primary source[2]Key Features
• Change · Context length Has a context window of 256k tokens.Compared with Mistral Small 3.2 ● Primary source[3]Context
• Change · Efficiency Mistral reports 40 percent less end-to-end completion time in a latency-optimized setup and three times more requests per second in a throughput-optimized setup than Small 3.Compared with Mistral Small 3 ● Primary source[1]

• Input text, image ● Primary source[3]Modalities [1]Key architectural details
• Output text ● Primary source[3]Modalities [2]Key Features
• Feature Reasoning modeConfigurable per request with the reasoning_effort parameter (none or high). ● Primary source[1]Reasoning on demand [2]Recommended Settings
• Feature Function calling ● Primary source[3]Features: Function Calling
• Open weights YesApache 2.0. ● Primary source[2] [1]
• Context window 256k tokens ● Primary source[3]Context [1]Key architectural details
• Parameters 119B total, 6.5B activeMixture of experts with 128 experts, 4 active per token. The announcement gives 6B active parameters per token (8B including embedding and output layers); the docs and the model card give 6.5B active. ● Primary source[3]description [2]Key Features
• Access API, Open-weights download, cloud partnerMistral API and AI Studio, Hugging Face, and NVIDIA (build.nvidia.com and NVIDIA NIM) at launch. ● Primary source[1]Availability

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

• Introducing Mistral Small 4Mistral AI · News · 16 March 2026

All model releases from Mistral AI on AI Radar →